ProofHelm is a zero-trust runtime AI security layer that sits between your apps/agents and any LLM provider. It enforces policies in real time and produces signed, audit-ready traces for compliance.
Built for modern AI systems: LLM chat endpoints, RAG assistants, multi-agent workflows, tool execution, and internal AI copilots.
Blocks jailbreaks, instruction override attempts, data exfiltration prompts, and policy bypass patterns.
Detects PII leakage, disallowed content, and hallucinated tool calls before they reach users or downstream systems.
Zero-trust enforcement for tool execution: allowlists, scopes, rate limits, and contextual rules per agent.
Works with any LLM provider (OpenAI, Azure OpenAI, Anthropic, Bedrock, local models) through a gateway or SDK.
Every decision ships with trace + cryptographic signature for audits and incident investigations.
Stream verdicts over WebSocket for dashboards, SOC alerts, and incident response workflows.
A layered pipeline minimizes latency and cost: fast local checks first, optional LLM guard as a last resort, and signed evidence for every step.
Output: { verdict, trace[], evidence_sig }
Same evidence model, same audit trail.
Start simple, then scale: self-hosted gateway, SDK middleware, or OpenAI-compatible proxy to cover every language with a base URL change.
Run ProofHelm as a FastAPI gateway and route LLM traffic through it.
POST /api/prompt → decision + evidence
POST /api/output → output firewall + evidence
WS /ws/firehose → live events
Integrate in your Python / Node services to attach rich context (agent name, tool, user, tenant).
Great for internal agents and microservices where you want fine-grained policy.
Use official SDKs — just change base_url to your ProofHelm gateway endpoint.
This approach is usually the fastest rollout for enterprises using mixed stacks.
ProofHelm is designed for security reviews and audit workflows. Every decision includes a detector trace and a cryptographic signature.
Prompt injection, data exfiltration, agent tool abuse, and output leakage are the new “runtime” risks in AI systems. ProofHelm’s approach is: cheap checks first, policy enforcement always, evidence on every decision.
A few quick answers for teams evaluating ProofHelm.
No. Provider safety filters are useful, but they’re not enough for enterprise controls like tool allowlists, audit trails, and org-specific policies. ProofHelm is an additional control layer.
The design is layered: most requests are decided by fast local checks. Optional LLM-based adjudication is budgeted and used only for ambiguous cases.
Yes. Run the gateway with Redis/Postgres (Docker Compose or Kubernetes). You can keep all traffic in your environment.
Yes. The policy layer is built for agent contexts: tool name, scopes, allowed domains/paths, and per-agent rules.
Start with a single gateway in front of one AI workflow. Turn on the WebSocket stream and review what gets blocked, then tighten rules and expand rollout.
Tell us what you’re building (RAG assistant, agent workflow, internal copilot) and we’ll show you how ProofHelm fits in.
Tip: Create a data/ folder with write permissions to store form submissions:
/data/leads.jsonl